How do you build a proactive agent?

Alknoma, in the proactive agent space, builds proactive AI agents for teams and tests them with Minutehand. A proactive agent needs four parts: a goal, a memory it writes as it works, a way to wake itself, and rules for when to act and when to wait. Minutehand then plays it through simulated days before it reaches real people.

Most agent tutorials build something reactive: a request comes in, the agent works, an answer goes out. A proactive agent turns that around. It starts itself, and most of the time the right thing for it to do is nothing.

1. Give it a goal it can finish

"Keep the project moving" never ends, so the agent never knows when to stop. "Confirm the offsite venue with Rosa and tell Owen" has an owner, a person to ask, and a done state. Give the goal a deadline too, so the agent knows when to give up and say so.

2. Give it a memory that outlives each wake

Between wakes the process may restart, so anything the agent needs next time must be written down. For each open item store who was asked, what, when, and when an answer is expected. On the next wake the agent reads its open items first, before it does anything else.

3. Give it ways to wake

• Events: a message in Slack or Teams, a ticket changing state, a webhook from a service it watches.
• Its own bookings: when it asks someone something, it books a wake for when the answer is due.
• A schedule: a daily or weekly pass, as a safety net rather than the main trigger.

Wakes it books for itself matter most. An agent that only runs on a fixed timetable either checks too often and nags, or too rarely and misses things.

4. Write down when to act, and when to wait

This is the part people skip and users notice. Make the rules explicit so they can be reviewed and tested:

• Do not follow up before the answer is due.
• Do not send the same message again with nothing new in it.
• When someone answers, come back to it promptly and pass the answer on.
• Never state a fact nobody gave you.
• At the deadline, stop and tell the owner where things stand.

5. Keep a person in the loop where it counts

Let the agent act alone on small, reversible things like a reminder, and ask a person before anything that is hard to undo or speaks for them in a new way. The line between the two belongs in the same rules as the timing.

6. Test it over time before it reaches people

A proactive agent's bugs are timing bugs, and they only show up over days. You cannot find them by sending one prompt and reading one reply. Play the agent through simulated days instead, with simulated colleagues who answer late or never, and check every move against your rules.

We built Minutehand for this step. It runs your agent against fake Slack, Teams, Jira, Asana, Notion, GitHub and Google Drive, owns the clock so days pass in seconds, and flags each follow-up sent too early or too late, each repeat, and each made-up fact, with what it was measured against.